
Summary
Written by a model from the transcript; not yet reviewed.
This episode explores the application of federated learning to agentic AI systems, featuring insights from Lorenzo Sani, a research scientist at Flower Labs and PhD student at Cambridge. The conversation covers how federated learning addresses privacy regulations by moving model training directly to decentralized data sources. Sani and the host examine how multi-step agent trajectories compound privacy risks and how collaborative learning enables enterprises to train on scattered data without exposing intellectual property or personal records. Additionally, the discussion touches on practical applications of open-source models, home automation projects, and the future of collaborative intelligence in sensitive domains like medical robotics.
- Federated learning ensures privacy by keeping raw data on edge devices and moving the training computations directly to the source. 9:06
- Agentic systems accumulate extensive private information across their long-running operational trajectories, including prompts and file system details. 20:24
- Open-source models offer greater control and transparency for privacy-sensitive applications compared to closed-source alternatives. 12:19
- Collaborative learning technologies enable organizations to aggregate insights from decentralized silos to improve model capabilities without centralizing raw data. 17:38
Chapters
- 0:02Podcast Introduction and Host Briefing
Chris Benson welcomes listeners, explains the episode’s focus on federated learning and agentic AI, and introduces Lorenzo Sani as the guest.
- 0:41Host and Guest Introduction
Chris introduces himself, notes Daniel’s absence, and formally welcomes Lorenzo Sani, a research scientist at Flower Labs.
- 5:47Federated Learning Fundamentals and Privacy
Lorenzo explains what federated learning is, its benefits, and how it preserves privacy by keeping data on edge devices.
- 8:51Privacy Guarantees and Differential Privacy Techniques
Discussion of statistical bounds, differential privacy, and challenges in ensuring models do not memorize sensitive data.
- 14:52Applying Federated Learning to Agentic Systems
Lorenzo describes how federated learning can train agentic AI models, using user trajectories and prompts while maintaining privacy.
- 30:42Practical Guidance for Deploying Federated Learning
Advice on starting with federated learning, collecting data, analyzing failures, and iterating to improve models in business contexts.
- 35:43Real‑World Use Cases and Trajectory Management
Examples of federated learning in medical data, language models, and agentic applications, highlighting trajectory complexity.
- 40:01Closing Thoughts and Takeaways
Chris wraps up, thanks Lorenzo, and encourages listeners to apply federated learning and agentic AI in their own projects.
- 40:56Future Vision for Federated and Agentic AI
Lorenzo reflects on the democratization of AI, potential for small‑scale deployments, and the role of privacy and interpretability.
- 42:57Outro and Call to Action
Hosts sign off, promote Practical AI resources, and thank partners and listeners.
Highlights
- 1:50
From Physics to Federated Learningstory
The guest explains the transition from studying physics in Bologna to researching privacy-preserving machine learning and federated learning applications.
- 4:37
Handling Distributed Data Complexityinsight
The discussion highlights how federated learning tackles the complexity of scattered and heterogeneous siloed data across edge devices.
- 7:48
Defining Federated Learning Mechanicsstrategy
Federated learning moves the training pipeline and all necessary computations to where the data is actually collected rather than a centralized data center.
- 9:48
Differential Privacy and Statistical Boundsinsight
Techniques like differential privacy provide statistical guarantees that trained models are secure and do not leak memorized private data.
- 16:04
Agentic Trajectories and Private Promptsinsight
Agentic systems introduce complex interaction trajectories containing sensitive user prompts, tool calls, and RAG systems that require careful privacy protection.
- 23:57
Halloween Party Agent Competitionstory
The guest shares a story about building an AI-powered platform for a Halloween costume competition where agents judged outfits and guests attempted prompt injection.
- 31:50
Scaling Down Model Requirementsdata
As AI models get smaller, cheaper, and easier to deploy, users can run advanced pipelines on local hardware for personal and enterprise projects.
Quotes
“I love the motivations behind it and the way to carry along the privacy‑preserving component to it.”
Unnamed speaker · 2:26 “Federated learning provides a good way to train on this distributed data that is very private because it contains your prompts.”
Unnamed speaker · 17:38 “Because whenever you put a model out of its comfort zone, so something that it's never tried before, it's going to make a lot of mistakes.”
Unnamed speaker · 37:54
Sponsored segments
- Hostinger 13:23–14:21
- Perplexity Computer 29:49–30:38
About this episode
How can AI models learn from private data without compromising privacy? Chris is joined by Lorenzo Sani, Research Scientist at Flower Labs and PhD researcher at Cambridge, to explore federated learning and privacy-preserving AI. They discuss how collaborative learning can expand access to private data for LLM training, how model trajectories can help improve agentic AI systems, and the benefits of open-weight models. The conversation also explores real-world applications of locally hosted AI agents, from healthcare and robotics to everyday automation (for a Halloween party). Featuring: Lorenzo Sani – LinkedIn https://www.linkedin.com/in/spencer-mcintyre-b48987194/Chris Benson – Website https://chrisbenson.com/, LinkedIn https://www.linkedin.com/in/chrisbenson, Bluesky https://bsky.app/profile/chrisbenson.bsky.social, GitHub https://github.com/chrisbenson, X https://x.com/chrisbensonLinks: Endeavor model page https://flower.ai/models/endeavorIntroducing Endeavor 1.1 https://flower.ai/blog/2026-11-03-introducing-endeavor-1.1Sponsors: Hostinger https://www.hostinger.com/in: Create your website today. Go to hostinger.com/practical and use practical for 10 percent off now. Perplexity AI https://www.perplexity.ai/gen/computer-lp/signin?_fallback_redirect=https%3A%2F%2Fwww.perplexity.ai%2Fgen%2Fcomputer-lp%2Fsignin%3Fpplx_account%3D49b12d57-abd5-43cd-8292-23ac5793257e&cl=p0yH9EGTSaSoYvCA&dub_id=p0yH9EGTSaSoYvCA&pcn=PRACTICAL&pcrn=EvergreenMultiModel&pplx_account=49b12d57-abd5-43cd-8292-23ac5793257e&pscn=PracticalAi&referrer=singular_click_id%3D8d3a8185-3295-4fa0-a0f2-1f2503087d8c&sl_id=j9oq&utm_campaign=PRACTICAL&utm_content=EvergreenMultiModel&utm_medium=PracticalAi&utm_source=Podcast&wpcn=PRACTICAL&wpcrn=EvergreenMultiModel&wpscn=PracticalAi&wpsrc=Podcast: "Start your free trial of Perplexity Computer today at pplx.ai/practical https://www.perplexity.ai/gen/computer-lp/signin?_fallback_redirect=https%3A%2F%2Fwww.perplexity.ai%2Fgen%2Fcomputer-lp%2Fsignin%3Fpplx_account%3D49b12d57-abd5-43cd-8292-23ac5793257e&cl=zN8r91RbQPOPm2s7&dub_id=zN8r91RbQPOPm2s7&pcn=PRACTICAL&pcrn=EvergreenMultiModel&pplx_account=49b12d57-abd5-43cd-8292-23ac5793257e&pscn=PracticalAi&referrer=singular_click_id%3D5c506b32-4d62-4f5b-9000-50a616249f55&sl_id=j9oq&utm_campaign=PRACTICAL&utm_content=EvergreenMultiModel&utm_medium=PracticalAi&utm_source=Podcast&wpcn=PRACTICAL&wpcrn=EvergreenMultiModel&wpscn=PracticalAi&wpsrc=Podcast and experience the next phase of AI. Resources and Events: Register for upcoming webinars here https://practicalai.fm/webinars! Prior Webinars from our partner Prediction Guard https://www.youtube.com/playlist?list=PLRpRdFmBfdF6boDgAD3nNaLm8Yx4KN2VtMidwest AI Summit 2026 https://midwestaisummit.com/Chris’s new book The Art of Swarming will be published soon, written to be the seminal foundational book for the fledgling swarming industry, defining the principles and architecture that enable autonomous agentic UxS swarms with emergent distributed embodied intelligence. You can find out more at his website https://chrisbenson.com/: https://chrisbenson.com
Transcript
7,254 words · 3 speakers
Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.
Topics
- Federated learning
- Agentic systems
- Flower Labs
- Cambridge
- LLMs
- Flower framework
- DeepMind
- MCP server
- RAG systems
- Prediction Guard
Page updated · transcript and summary are machine-generated